search

luckiezhou / DynamicTriad

Dynamic Network Embedding by Modeling Triadic Closure Process
Apache License 2.0
138 stars 49 forks source link

About academic dataset #11

Open rootlu opened 5 years ago

rootlu commented 5 years ago

I analyzed the published data set, the number of static edges is only 714470, which is inconsistent with the paper, is it different from the used dataset in the paper?

luckiezhou commented 5 years ago

The number of edges reported in the paper is not counted on static graphs, but counted on the graphs built under our time step configuration. 16 graphs are generated from the original static graphs in our paper, and each of them is the combination (summing up the weights for the same edge) of 4 consecutive static graphs. Specifically, the 16 graphs are built from static graphs 0-3, 2-5, 4-7, ..., 30-33. Using our script, the sequence of graphs are generated by setting "-l 4 -s 2 -n 16".

skx300 commented 5 years ago

Hi Zhou, why there is an overlapping between each time step? While for the other two datasets, Mobile and Loan, there is no overlapping.

luckiezhou commented 5 years ago

It is not guaranteed that each researcher, especially graduate students, publishes at least one paper per year. In order to build a more temporally stable dynamic network, I have to merge the records in a short period (4 years considering the average time a Ph.D. student focuses on publications) to build the graph of a single timestep. I think the ideal value for the stride is 1 (i.e. a 3-year-overlap) so that each student-mentor link may appear in several consecutive time steps, however, I choose 2 simply because there are a moderate number of time steps generated under this setting.

skx300 commented 5 years ago

Thank you Zhou. Does it mean all vertices should appear in every time snapshot?

luckiezhou commented 5 years ago

Yes. All the snapshots share the same vertex set, although some vertices might be isolated in certain snapshots.